[KAFKA-8253] Unrecoverable state when a broker's INPUT access is blocked (Zombie broker)

• Type: Bug
• Status: Open
• Priority: Critical
• Resolution: Unresolved
• Affects Version/s: 2.1.1
• Assignee: Unassigned
• Reporter: Gowtham Gutha
• Created: 18/Apr/19 06:11
• Updated: 18/Apr/19 06:11

Description

I carried out a test as mentioned in my SO question

https://stackoverflow.com/questions/55706589/what-happens-if-the-leader-is-not-dead-but-unable-to-receive-messages-in-kafka

Gist of the test:

A broker's INPUT access is blocked. So it is not able to receive any messages.

But still it can send heartbeats to ZK, so that a leader election will not happen.

So any message produced to the partition lead by this zombie broker is never produced leaving the system in an unrecoverable state.

 

Possible resolution:

There should be a 2 way communication such that, if a broker is not able to have any INPUT access, the ZK MUST know of it by sending some ping messages to the brokers.

If there is no response from the broker, elect a new one. Since, if the broker is not ping-able by ZK, that broker is as good as dead for its purpose.